TDTrack: Transformer Tracking via Target Cues Generator and Distractor Handling Network

Benefiting from Transformer's powerful global modeling capabilities, many Transformer-based trackers have emerged. However, trackers using fixed templates usually have difficulty capturing changes in target appearance, and the introduction of multiple templates brings high computational complexity. In addition, most trackers ignore the handling of distractors and may fail to track when similar objects appear in the background. To solve the above problems, this paper proposes a Transformer tracking algorithm based on a target cue generator and a distractor handling network. Specifically, a target cue generator is first designed to use very few target cue tokens instead of multi-frame template images, which reduces computational complexity while capturing changes in target appearance. In addition, a distractor handling network is designed to enhance the tracker's ability to discriminate distractors through simple post-processing operations. Extensive experiments show that TDTrack achieves high tracking accuracy on four challenging benchmark datasets and runs in real-time at 78.5 FPS.